Om P. Damani

ONE-IP: techniques for hosting a service on a cluster of machines

Abstract With the explosive growth of the World Wide Web, some popular web sites are getting thousands of hits per second. As a result, clients (browsers) experience slow response times and sometimes may not be able to access some web sites at all. Upgrading the server nodes to more powerful machines may not always be cost-effective. A natural solution is to deploy a set of machines, or a cluster, and have them work together to host a single service. Such a server cluster should preferably publicize only one server name for the entire cluster so that any configuration change inside the cluster does not affect client applications. In this paper, we first discuss existing approaches to distributing clients requests for a single service to different machines in a cluster. We then propose two new techniques, collectively called ONE-IP , based on dispatching packets at the IP level. They have the advantages of fast dispatching, and ease of implementation. Ideas presented here are generic and should be applicable to other services as well.

Distributed recovery with K-optimistic logging

Fault-tolerance techniques based on checkpointing and message logging have been increasingly used in real-world applications to reduce service downtime. Most industrial applications have chosen pessimistic logging because it allows fast and localized recovery. The price that they must pay, however, is the higher failure-free overhead. In this paper, we introduce the concept of K-optimistic logging where K is the degree of optimism that can be used to fine-tune the tradeoff between failure-free overhead and recovery efficiency. Traditional pessimistic logging and optimistic logging then become the two extremes in the entire spectrum spanned by K-optimistic logging. Our approach is to prove that only dependencies on those states that may be lost upon a failure need to be tracked on-line, and so transitive dependency tracking can be performed with a variable-size vector. The size of the vector piggybacked on a message then indicates the number of processes whose failures may revoke the message, and K corresponds to the system-imposed upper bound on the vector size.

How to recover efficiently and asynchronously when optimism fails

We propose a new algorithm for recovering asynchronously from failures in a distributed computation. Our algorithm is based on two novel concepts-a fault-tolerant vector clock to maintain causality information in spite of failures, and a history mechanism to detect orphan states and obsolete messages. These two mechanisms together with checkpointing and message-logging are used to restore the system to a consistent state after a failure of one or more processes. Our algorithm is completely asynchronous. It handles multiple failures, does not assume any message ordering, causes the minimum amount of rollback and restores the maximum recoverable state with low overhead. Earlier optimistic protocols lack one or more of the above properties.

Fault-tolerant distributed simulation

In traditional distributed simulation schemes, the entire simulation needs to be restarted if any of the participating logical processes (LPs) crash. This is highly undesirable for long running simulations. Some form of fault tolerance is required to minimize the wasted computation. A rollback based optimistic fault tolerance scheme is integrated with an optimistic distributed simulation scheme. In rollback recovery schemes, checkpoints are periodically saved on stable storage. After a crash, these saved checkpoints are used to restart the computation. We make use of the novel insight that a failure can be modeled as a straggler event with the receive time equal to the virtual time of the last checkpoint saved on stable storage. This results in saving of implementation efforts, as well as reduced overheads. We define stable global virtual time (SGVT), as the virtual time such that no state with a lower timestamp will ever be rolled back despite crash failures. A simple change is made in existing GVT algorithms to compute SGVT. Our use of transitive dependency tracking eliminates antimessages. LPs are clubbed in clusters to minimize stable storage access time.

Reliability and availability issues in distributed component object model (DCOM)

Distributed Component Object Model (DCOM) is one of the emerging standards for distributed objects. Before DCOM can be used to build mission-critical applications, the reliability and availability issues must be addressed. In this position paper, we outline the current research directions of the InterCOM project, which exploits the dynamic behavior, the extensible architecture, and the component software model of DCOM to provide fault tolerance capabilities to distributed applications.

Transliteration for Resource-Scarce Languages

Today, parallel corpus-based systems dominate the transliteration landscape. But the resource-scarce languages do not enjoy the luxury of large parallel transliteration corpus. For these languages, rule-based transliteration is the only viable option. In this article, we show that by properly harnessing the monolingual resources in conjunction with manually created rule base, one can achieve reasonable transliteration performance. We achieve this performance by exploiting the power of Character Sequence Modeling (CSM), which requires only monolingual resources. We present the results of our rule-based system for Hindi to English, English to Hindi, and Persian to English transliteration tasks. We also perform extrinsic evaluation of transliteration systems in the context of Cross Lingual Information Retrieval. Another important contribution of our work is to explain the widely varying accuracy numbers reported in transliteration literature, in terms of the entropy of the language pairs and the datasets involved.

Experiences with English-Hindi, English-Tamil and English-Kannada Transliteration Tasks at NEWS 2009

We use a Phrase-Based Statistical Machine Translation approach to Transliteration where the words are replaced by characters and sentences by words. We employ the standard SMT tools like GIZA++ for learning alignments and Moses for learning the phrase tables and decoding. Besides tuning the standard SMT parameters, we focus on tuning the Character Sequence Model (CSM) related parameters like order of the CSM, weight assigned to CSM during decoding and corpus used for CSM estimation. Our results show that paying sufficient attention to CSM pays off in terms of increased transliteration accuracies.

Optimistic recovery in multi-threaded distributed systems

The problem of recovering distributed systems from crash failures has been widely studied in the context of traditional non-threaded processes. However, extending those solutions to the multi-threaded scenario presents new problems. We identify and address these problems for optimistic logging protocols. There are two natural extension to optimistic logging protocols in the multi-threaded scenario. The first extension is process-centric, where the points of internal non-determinism caused by threads are logged. The second extension is thread-centric, where each thread is treated as a separate process. The process-centric approach suffers from false causality while the thread-centric approach suffers from high causality tracking overhead. By observing that the granularity of failures can be different from the granularity of rollbacks, we design a new balanced approach which incurs low causality tracking overhead and also eliminates false causality.

Interference-constrained coverage algorithms in the protocol and SINR models

AbstractA wireless network’s design must include the optimization of the area of coverage of its wireless transmitters—mobile and base stations in cellular networks, wireless access points in WLANs, or nodes on a transmit schedule in a wireless ad-hoc network. Furthermore, with increasing densities of wireless network deployments, paucity of spectrum, and new developments like whitespace devices and cognitive networks, there is a need to study the computational efficiency of managing interference and optimizing coverage. This work presents new algorithms for computing and optimizing interference-limited coverage of wireless networks under protocol and Signal-to-Interference-and-Noise Ratio (SINR) models. For the protocol model we demonstrate lower bounds on computation of the coverage area for an